Friendli Suite Inference Platform

FriendliAI Corp is an AI inference company headquartered in San Francisco with a hub in Seoul, led by CEO and co-founder Byung-Gon Chun. It addresses the high cost and latency of running large language and multimodal models in production, where most operational time is spent on inference.

Friendli Suite lets teams call open models through Model APIs, deploy their own or Hugging Face models on Dedicated Endpoints with autoscaling, or bring their own GPUs. Its engine uses custom GPU kernels, smart caching, continuous batching, speculative decoding and parallel inference, and the company claims more than twice the speed of standard serving with 99.99% uptime SLAs. It offers day-zero access to models such as NVIDIA Nemotron and GLM. Customers include LG AI Research, SK Telecom, Upstage, Liner and Scatter Lab.

• Model APIs: pay-as-you-go access to open models with day-zero support for new releases.
• Dedicated Endpoints: autoscaling deployments of custom or Hugging Face models on reserved GPUs.
• Bring Your Own GPU: runs the Friendli engine on customer-owned GPU infrastructure.
• Inference optimizations: custom kernels, caching, continuous batching and speculative decoding.
• InferenceSense: lets GPU cloud operators monetize idle capacity.
• Enterprise controls: SOC 2 Type II and HIPAA compliance with 99.99% uptime SLAs.

Vietnamese chatbot and call-center vendors can serve fine-tuned open models at lower GPU cost. Telecom or cloud providers in Vietnam with GPU clusters can run the engine on their own hardware to offer inference APIs. Software firms can power coding or document agents with open models through one API.

The company states its engine cuts GPU costs by up to 90% while preserving model accuracy. Liner reports about 30% lower LLM cost per turn and NextDay AI scaled to trillions of tokens with 50% fewer GPUs. One-click deployment of hundreds of thousands of Hugging Face models shortens time to production; LG AI Research had a custom model API live in about a day.

As generative AI moves from pilots to production and agentic workloads, inference has become the largest share of AI operating cost. Enterprises increasingly adopt open-weight models to control cost and data, creating demand for optimized serving platforms, including in Asian markets building local-language models.

FriendliAI Corp is based in San Francisco with a hub in Seoul; it opened a larger San Francisco office in 2026. Raised a 20 million USD seed extension in August 2025 led by Capstone Partners with Sierra Ventures, Alumni Ventures, KDB and KB Securities. Partners include LG AI Research (EXAONE), NVIDIA, Z.ai, Hugging Face and Ollama.